Market Brief: Friday, July 17, 2026

Today's lesson

Why Reasoning Models Cost More to Serve

Reasoning models generate long chains of thought before answering, multiplying output tokens — and output tokens drive inference cost. Why it matters today: Today's lesson connects the Model Benchmarks & AI Compute Economics track to the compute market readers are watching.

Term of the Day

Cloud quota

A cloud quota is the limit a provider places on how much capacity a customer can use. Even if a GPU type appears in a cloud catalog, a customer may not be able to access many of them without quota approval. Quotas are especially important for scarce AI accelerators.

H100 IndexiWhat it costs to rent one H100 GPU for an hour on demand — the market’s anchor price.

Value

$4.12 / H100-hr

Last Price Change

Sep 16, 2026

Spot Compute PriceiThe cheaper, interruptible GPU price — a read on whether spare capacity is available.

Value

$2.46 / H100e-hourH100e-hour means an H100-equivalent GPU hour, a normalized unit used to compare different GPUs and providers.

Last Price Change

May 27, 2026

Model Serving CostiRoughly what it costs to run a model and generate its output at scale, per million tokens.

Value

$2.78 / 1M tokens

Last Price Change

Aug 8, 2026

GPU AvailabilityiHow easy or hard it is to actually get GPU capacity right now, scored 0–100.

Value

78 / 100 score

Last change

Sep 17, 2026

Jul 16, 2026

Nvidia partners with Noetra for 140MW Vera Rubin cluster in Japan

The proposed 140 MW Vera Rubin cluster would add future accelerator capacity in Japan, but power delivery, construction, and hardware deployment mean it does not increase immediately available GPU supply.

Jul 16, 2026

Australia's Macquarie DC acquires 34,200 sqm site in Sydney for 200MW data center campus

Acquiring the Sydney site creates an option for a 200 MW campus, but land ownership is an early development milestone rather than operational data-center or GPU capacity.

Jul 16, 2026

Japan Government, Industrial Leaders and NVIDIA Launch the World’s First National AI Infrastructure

Government-backed national AI infrastructure can expand domestic accelerator access and procurement demand, but the capacity effect depends on funded deployments becoming operational.

Jul 15, 2026

The grid’s fastest-growing resource isn’t generation. It’s flexibility.

More flexible grid demand can help utilities integrate large data-center loads without matching every peak with new generation, though the article alone does not establish lower power or compute prices.

Jul 15, 2026

PJM capacity prices hit price cap, reserve shortfall grows

A capacity auction at the price cap with a reserve shortfall raises the cost and reliability risk of adding large loads in PJM, pressuring data-center power economics before it changes GPU list prices.

H100 GPU Pricing Report: August 2026

Canonical CT-H100 median, reviewed H100 provider rows, freshness dates, and composition caveats in one citable report.

Previous issue: Market Brief: Thursday, July 16, 2026

Virginia SCC weighs Dominion data center transmission cost allocation

Next issue: Market Brief: Saturday, July 18, 2026

NVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads — a Key Metric for Agentic AI